My name is Sukrit and I am an AI/ML Algorithm Engineer at KLA working on researching, prototyping,
and deploying cutting-edge deep learning models for computer vision tasks on electron microscope images
of semiconductor chips. My work and interests include unsupervised + self-supervised learning for anomaly
detection, generative modeling, deep metric learning, and explainable AI.
Before my time at KLA, I was getting my MS from UC Berkeley, where I researched and developed deep learning
applications for medical image (MRI) denoising, reconstruction, deblurring, and stitching, which you can read
more about here.
During my time at Berkeley, I also enjoyed TA’ing several different courses, from computer architecture to signals
and optimization theory. Read more about my teaching work here. I still teach and mentor when I can in my
current role at KLA. I recently co-taught an internal Introduction to Deep Learning
course with Prof. Yoav Ram, and regularly mentor summer interns.
Check out some of the projects that I have worked on in the past here
During my summer at Apple, I worked on two separate projects. In both, my work entailed applying software engineering principles in a cybersecurity context. All my work is under an NDA, but I have a brief description of both projects below.
My first project was a solo project, in which I created a Command Line Interface in order to perform complex queries on a distributed graph database.
My second project was a group project, where I designed and implemented a server automation workflow. My responsibilities specifically were related to backend scripts and server orchestration.
Proposed a modification of the dictionary-based Sparse Representation-based Classification (SRC) facial recognition algorithm using different low dimensional representations of input images. The methods of low dimensional representation proposed included: naive decimation, filtered downsampling, Haar Wavelet downsampling, Principle Component Analysis, and a learned Convolutional Autoencoder.
Analyzed UC Berkeley enrollment data from several perspectives: Built a variety of both binary classifiers and regression models to predict the location of a school and admitted white male proportion, respectively, given multiple gender and ethnic features. Additionally, interpreted enrollment data from a Fairness in ML perspective.
Face Morphing Project shown
Several projects, click on each link below for more in depth:
Made a RISC-V 60 MHz CPU with a 3-stage pipeline that accounts for data and control hazards. The CPU is also connected to a user I/O interface and synthesizer using memory mapped I/O. The synthesizer is a monophonic subtractive synthesizer that is constructed using a numerically controlled oscillator and is capable of producing four types of tones (sinusoidal, sawtooth, square, and triangular).
Made a Crazyflie drone improviser randomly navigate between three waypoints while avoiding a Kobuki turtlebot adversary. Control logic dictated by Reactive Control Improvisation algorithm, which creates an FSM using specified LTL constraints. Localization done using Optitrack Motion Capture, and orchestration done via ROS.
Constructed a deep learning pipeline for automatic image object removal. This was done by combining the Mask R-CNN network, which does object recognition, with a Generative Image Inpainting network, which uses a multiscale contextual GAN to inpaint specified regions.
3rd in Class
Transmitted an image across the room using a radio and custom image compression scheme. For transmission, we implemented our own Audio Frequency Shift Keying (AFSK) modem that would modulate and demodulate the compressed data to and from the radio, respectively. For the image compression, we impelemented our own variant of JPEG image compression from the ground up, that used a combination of color channel tranformation, a Haar wavelet transform, quantization, and LZW lossless compression.
Created a self-inverting pendulum by combining a LQR Controller with a Leunberger Observer for state estimation and Bang-Bang control for swing-up.
Generated new music using a random walk on a Markov Chain. Transition probablilities of the Markov Chain were created by parsing and analyzing a variety of MIDI files of different genres.
Implemented a network router simulator using distance-vector routing, a distributed routing algorithm. Additionally, built a Wide Area Network (WAN) optimizer, a middlebox that optimizes the amount of data transmitted over a wide area network to reduce excessive bandwidth costs.
Created a musical instrument in Max8 that generates music based on keyboard inputs and movements tracked via a webcam using a Canny Edge Detection filter for processing.
3rd Place, Hack UCSC
Food Analyzer is a Chrome extension that uses web scrapers, machine learning, and natural language processing to instantaneously obtain relevant nutritional information for recipes and ingredients seen online. It allows the user to identify unhealthy ingredients and make alternative decisions.
3rd Place, Robotics at Berkeley Dorm Ex Machina Competition
As a part of the Dorm Ex Machina Competition, we created a Automated Snack Delivery Robot. We CAD’ed and built the robot from the ground up with a $100 budget using a Raspberry Pi, OpenCV, Homemade Line Sensors, and a Motor Controller
Created a wallpaper setter in python that used Reddit APIs to get the latest landscape pictures from r/ EarthPorn and set them as my desktop background. Additionally, I wrote a plist script in order to automate image getting
CalHacks 3.0
Created MileStones, a scavenger hunt web app that takes users from one location to the next based on user feedback, allowing them to discover and explore new places that match their taste. Used Yelp, Weather.com, and IBM Watson APIs and Flask Libraries to create a working prototype in only 36 hours time
1 st Place BattleBots Competition Winner
Competitive residential summer program at UPenn in Robotics
Food Watchers is an iPhone application designed in partnership with Stanford’s Lucile Packard Hospital’s Pediatric Weight Management Program. It encourages healthy food habits for obese participants in a simple and accessible way. It has been rolled out to current participants of the program and has over 1000 downloads in 5 different continents
HackingEDU Hackathon
Developed a beta version of a sign language to text converter. Developed on a LEAP motion with Python using LEAP motion API and a neural network machine learning algorithm to make hand recognition faster and more accurate
